Skip to content

Latest commit

 

History

History
63 lines (51 loc) · 2.78 KB

eval.md

File metadata and controls

63 lines (51 loc) · 2.78 KB
Raw

NeRS Evaluation

Evaluation Data

Download the evaluation data from here:

gdown https://drive.google.com/uc?id=1Nle6efI8HAHIXgvqhBhEb7ZHRn2dg5Mm
unzip -n ners_evaluation.zip -d data

The metadata for the evaluation sequences contains an extra set of camera parameters (camera_pretrained). In total, we provide 3 types of cameras:

  • camera_initial: the off-the-shelf camera using PoseFromShape
  • camera_pretrained: the camera parameters after optimizing camera_initial to minimize silhouette reprojection error using the template car mesh (i.e. the cameras after Stage 1 in the paper). This camera serves as reasonably well-aligned but inexact camera initialization.
  • camera_optimized: the camera parameters after optimizing NeRS + some manual tuning. These cameras should be better than the ones in the main dataset. I manually fixed some of the rear viewpoints, which was a common failure mode for PoseFromShape.

We consider two evaluation settings:

  1. Novel view synthesis with fixed cameras (camera_optimized) treats each of the N - 1 input cameras as fixed. At inference time, given the Nth target camera, we render the Nth target image and compare this with the ground truth image.
  2. In-the-wild novel view sysnthesis (cameras_pretrained) treats each of the N - 1 input cameras as a noisy initialization. During training, the cameras are also optimized. At inference time, given a target view and possibly given a noisy target camera, we search for the camera that best aligns with the target view.

The evaluation data contains the images corresponding to the NeRS renderings from both of these evaluation settings. These can be evaluated using:

$  python -m eval.eval --pred-name ners_fixed
Name            MSE   PSNR   SSIM  LPIPS
ners_fixed   0.0254   16.5  0.720  0.172
$  python -m eval.eval --pred-name ners_trained
Name            MSE   PSNR   SSIM  LPIPS
ners_trained 0.0364   15.0  0.673  0.206

You can also retrain all models on the evaluation data. For example, to retrain the fixed camera evaluation:

python -m eval.train_evaluation_model --data_dir data/evaluation --fix-cameras \
    --instance-id 7251103879 --camera-index 0 --camera-type camera_optimized

and the trainable camera evaluation:

python -m eval.train_evaluation_model --data_dir data/evaluation \
    --instance-id 7251103879 --camera-index 0 --camera-type camera_pretrained

Note that one model must be trained for each camera index for each instance. This process can be automated by running:

python -m eval.eval_driver --data-dir data/evaluation --evaluation-mode fixed

python -m eval.eval_driver --data-dir data/evaluation --evaluation-mode in-the-wild